GTC Berlin recap: Huang sells Europe a full AI stack and calls agent security job one

At GTC Berlin, Nvidia’s chief executive framed Europe as ready to produce AI rather than consume it, capping a week in which the company says it committed nearly $10 billion to Europe’s technology ecosystem and works with more than 12,000 partners. The keynote at Berlin’s Tempodrom, part of a three-day event with more than 100 sessions, covered Vera Rubin in full production, the modular DSX POD reference design, open models, an agent knowledge layer (Agent Knowledge Foundation), telecom, healthcare, cars, a reference humanoid and RTX Spark laptops (local AI lineup). Other announcements included a first commercial phone call on T-Mobile’s network with Nokia, wider adoption of the Hyperion automotive platform by Volvo Cars and others, and the Agile ONE SN reference humanoid expected next year. Hugging Face’s Microduck robots also drew attention (Microduck preorders).

Much of the keynote focused on agents and containing them. Huang described an agent as five parts: a model, a harness that orchestrates work and memory, a sandbox that isolates the harness, knowledge and tools. He argued agent privileges should be minimal, monitored and adjusted each time an agent launches. Nvidia’s product answer is the OpenShell Reference Architecture, an open-source, policy-controlled runtime with kernel-level sandboxing between an agent and the systems it can affect, backed by cloud partners (cloud partners). Naming least privilege as the default sets an agenda for enterprise buyers; independent security reviews of the sandboxing will show whether it holds in practice.

Nvidia says DSX POD, prebuilt Vera Rubin AI factories in modular containers, can shrink onsite deployment from six months to 14 days once site utilities are ready, and that its DSX AI Factory Platform can cut cooling water use by up to 100% and deliver the same output with up to 30% less power. These are vendor claims with caveats built in, notably the utilities condition; power and grid access, more than containers, slow European data centers (plug-in data centers).

Huang urged companies to post-train open models like Nemotron on their own data, calling the result ‘sovereign intelligence.’ Open models with reproducible data and scripts are useful, but sovereignty on one American company’s chips, libraries and reference designs is narrower than the word suggests. The applied examples were concrete: agents at Lila Sciences helping design a potential B-cell lymphoma therapy, Siemens testing production plans in a digital twin, and adaptive radiation therapy with Siemens Healthineers that tracks a changing tumor during treatment. Real-time radiation targeting in particular requires clinical validation and regulatory review well beyond a keynote demonstration. More in our Nvidia coverage, AI agents coverage and robotics coverage.